Aims: Photoplethysmography (PPG) is a non-invasive and low-cost method widely used in wearable health monitoring devices for tracking cardiovascular activity. Its reliability is often limited by motion artifacts, low perfusion, and inter-subject variability, particularly in real-world applications. Accurate heartbeat detection from PPG without relying on external references such as ECG is essential. This study presents the Peakwise Correlation Pulse Detector (PCPD), a novel algorithm designed for robust and stand-alone systolic peak detection in PPG signals under noisy conditions.
Methods: The PCPD algorithm integrated five key steps: (1) signal segmentation into overlapping windows, (2) cross-correlation with ideal waveform templates derived from high-quality beats, (3) extraction of a Minimum Correlation Curve (MCC) for peak enhancement, (4) morphological validation via a quadratic linear classifier using skewness and the Ratio of Area (RoA) as features, and (5) physiological constraint checks for final verification. The method was evaluated on two datasets: a custom 12-hour high-quality PPG set and the BIDMC dataset. To simulate noise, artificial pulse-free segments were added. The final evaluation comprised 139,953 labeled peaks, with 47% true positives and 53% true negatives.
Results: Under clean signal conditions, PCPD achieved 99.13% accuracy on the custom dataset and 97.04% on BIDMC. Under noisy conditions, the method maintained high accuracy with 99.30% on the proposed dataset and 97.16% on BIDMC. Precision, recall, and F1 scores also remained consistently high across all conditions. In a comparative evaluation involving eleven benchmark peak detection algorithms, PCPD demonstrated superior performance, especially under noisy conditions.
Conclusions: The PCPD algorithm has demonstrated high accuracy and strong noise resilience. Its integration of morphological analysis and physiological constraints makes it a reliable solution for real-time heartbeat detection and signal quality assessment in wearable PPG-based monitoring systems in real-world environments.